Pre-Action Summary and Confirmation Prompt Template

This prompt is designed for agent builders and platform engineers who need to create a reliable, human-readable checkpoint before an AI system executes an action with side effects. The core job-to-be-done is to pause an autonomous workflow and generate a structured summary that captures the intent, scope, risk level, and explicit rollback plan for a proposed action. The ideal user is a developer integrating this prompt into an agent harness where a human operator or an approval service must review and confirm the action before execution proceeds. This is not a prompt for generating creative text or answering questions; it is a control-plane component for high-stakes operations such as database writes, API mutations, financial transactions, or infrastructure changes.

You should use this prompt when the cost of an incorrect autonomous action is high and a human-readable audit trail is required. It is appropriate for workflows where the action type, target resource, and implications can be clearly defined but the final decision requires human judgment. Do not use this prompt for low-risk, idempotent, or fully reversible read-only operations where a confirmation step adds unacceptable latency. It is also unsuitable for real-time, sub-second decision loops where a human cannot reasonably review the summary in time. The prompt assumes that the calling application has already resolved the action details and can provide them as structured input; it does not perform intent disambiguation or gather missing information from the user.

Before implementing this prompt, ensure your harness can supply the required variables: the action type, the target resource identifier, a clear statement of the intended change, a risk level classification, and a defined rollback or revert plan. If any of these inputs are missing or uncertain, pair this prompt with an Ambiguous Intent Clarification or Missing Information Interruption prompt first. The output of this prompt should be treated as a blocking gate: the system must not proceed to execution until a confirmed approval signal is received. For regulated domains, always log the generated summary, the approval decision, and the reviewer's identity as part of your compliance evidence package.

This template is the core of the pre-action confirmation workflow. It instructs the model to produce a structured checkpoint that captures the intent, scope, risk, and required confirmation for any planned action. The output is designed to be consumed by a human reviewer or an upstream orchestration system before an irreversible step is taken. Use this template as a starting point and adapt the placeholders to your specific domain, whether that's a database write, a financial transaction, or a configuration change.

textCopy
You are an AI agent preparing to execute a planned action. Before proceeding, you must generate a structured summary for human review and confirmation. Do not execute the action. Only generate the summary.

## Action to Summarize
[ACTION_DESCRIPTION]

## Action Type
[ACTION_TYPE]

## Target Resource
[TARGET_RESOURCE]

## Context
[CONTEXT]

## Constraints
[CONSTRAINTS]

## Output Schema
Generate a JSON object with the following fields:
- "action_id": A unique identifier for this action.
- "summary": A one-sentence, plain-English summary of what will happen.
- "intent": The business or operational goal this action serves.
- "scope": A detailed description of what is affected, including specific resources, systems, or data.
- "risk_level": One of "low", "medium", "high", "critical".
- "risk_rationale": A brief explanation for the assigned risk level.
- "irreversible": A boolean indicating if the action can be easily undone.
- "rollback_plan": If irreversible is false, describe the steps to undo the action. If true, state "No automated rollback is possible."
- "expected_outcome": A description of the system state after successful execution.
- "confirmation_required": A boolean, set to true.
- "confirmation_prompt": A direct question to the human reviewer, asking for explicit approval to proceed.

## Example Output
{
  "action_id": "deploy-v2.3.1-canary",
  "summary": "Promote service 'checkout-api' version v2.3.1 to the canary stage, shifting 5% of production traffic.",
  "intent": "Gradually roll out a new feature flag and performance improvements with minimal blast radius.",
  "scope": "The 'checkout-api' deployment in the 'production' cluster, canary namespace. 5% of traffic from the 'web' load balancer will be routed to the new version.",
  "risk_level": "medium",
  "risk_rationale": "The change is limited to 5% of traffic and the canary stage has automatic rollback triggers on error rate increase. However, it is a production change.",
  "irreversible": false,
  "rollback_plan": "Execute the CI/CD pipeline's 'rollback-canary' job, which will shift all traffic back to the stable version v2.3.0.",
  "expected_outcome": "5% of users will be served by checkout-api v2.3.1. Monitoring dashboards will show a new version in the canary segment.",
  "confirmation_required": true,
  "confirmation_prompt": "Proceed with promoting checkout-api v2.3.1 to the canary stage? (yes/no)"
}

To adapt this template, replace the square-bracket placeholders with the specific details of the action your agent is about to take. The [ACTION_DESCRIPTION] should be a raw, unstructured description from the planning step. The [ACTION_TYPE] and [TARGET_RESOURCE] help the model categorize the risk. The [CONSTRAINTS] block is where you inject operational policies, like 'All production changes require a rollback plan' or 'Actions on PII data require a data minimization statement.' After generating the summary, your application harness should parse the JSON and present the confirmation_prompt to the designated human reviewer. Do not proceed with the action unless the reviewer explicitly approves.

The Pre-Action Summary and Confirmation prompt is not a standalone chat interaction; it is a synchronous checkpoint inside an agent execution loop. The application must intercept the agent's intent to perform a write, mutation, or irreversible operation, call the LLM with this prompt to generate the summary, and then block further execution until a human reviewer provides explicit confirmation. The harness is responsible for providing the structured [ACTION_DETAILS] context, enforcing the output schema, and routing the generated summary to the correct approval queue.

To implement this, wrap the prompt in a function that is called by the agent's tool-execution middleware. The function should accept a dictionary of action details (type, target, payload, risk level, rollback plan) and format them into the [ACTION_DETAILS] placeholder. After calling the LLM, validate the response against a strict JSON schema that requires summary, risk_level, affected_resources, rollback_plan, and requires_confirmation fields. If validation fails, retry once with the validation error injected into the [CONSTRAINTS] block. If it fails again, log the raw output and escalate to a human operator with a PRE_ACTION_SUMMARY_FAILURE alert. The model choice here should prioritize instruction-following and schema reliability (e.g., GPT-4o, Claude 3.5 Sonnet) over speed, as this is a synchronous blocking step.

Once a valid summary is generated, the harness must render it in a human-readable format (Markdown or a UI card) and push it to a review queue. The queue item should include a unique action_id, the full summary, a timestamp, and two buttons: Approve and Reject. The agent's execution thread must block on this step, polling for a decision or waiting for a webhook callback. A timeout should be set (e.g., 15 minutes for standard operations, 5 minutes for incident response). If the timeout expires, the harness must treat it as a rejection, log the event, and execute the rollback plan if one was provided. Do not proceed with the action on a timeout. For high-risk actions defined by a risk_level of critical or high, require multi-stakeholder approval by routing the summary to a designated approval chain before the agent can proceed.